English

Linear Convergence of a Frank-Wolfe Type Algorithm over Trace-Norm Balls

Machine Learning 2017-11-10 v3 Data Structures and Algorithms Optimization and Control Machine Learning

Abstract

We propose a rank-kk variant of the classical Frank-Wolfe algorithm to solve convex optimization over a trace-norm ball. Our algorithm replaces the top singular-vector computation (11-SVD) in Frank-Wolfe with a top-kk singular-vector computation (kk-SVD), which can be done by repeatedly applying 11-SVD kk times. Alternatively, our algorithm can be viewed as a rank-kk restricted version of projected gradient descent. We show that our algorithm has a linear convergence rate when the objective function is smooth and strongly convex, and the optimal solution has rank at most kk. This improves the convergence rate and the total time complexity of the Frank-Wolfe method and its variants.

Keywords

Cite

@article{arxiv.1708.02105,
  title  = {Linear Convergence of a Frank-Wolfe Type Algorithm over Trace-Norm Balls},
  author = {Zeyuan Allen-Zhu and Elad Hazan and Wei Hu and Yuanzhi Li},
  journal= {arXiv preprint arXiv:1708.02105},
  year   = {2017}
}

Comments

In NIPS 2017

R2 v1 2026-06-22T21:08:35.102Z